A ^double/^long return hint on a fn's name now (a) coerces the fn's value on the
way out — exact->inexact / jolt->fx, like a JVM primitive return — and (b) types a
call to it, so an accumulator over the result specializes:
(defn ^double work [^double x ^double y] (+ (* x x) (* y y)))
(loop [acc 0.0] (recur (+ acc (work a b)))) ; (+ acc (work ..)) -> fl+
The analyzer pushes the name's numeric tag onto each arity (:ret-nhint) for the
back-end coercion, and resolve-global surfaces the callee's declared return
(:num-ret, read from var meta) onto the :var node so jolt.passes.numeric types the
call. defn carries the name hint through.
This unblocks the accumulator-over-fn-result pattern that round 2 had to demote.
The win is bounded by call overhead in an open/dispatched build (~1.15x on a hot
loop whose body is a helper call); it compounds with direct-linking and, later,
inlining. A numeric return hint is a contract, like ^long — redefining the var to
return another type in an open build breaks it.
Not yet: per-arity arglist return hints, (defn f (^double [x] ..)). Gate:
test/chez/numeric-test.ss 39/39; full make test green, 0 new corpus divergences.
Wire the optimization gate to build mode. inline-enabled? (which gates the
inference + flatten-lets + scalar-replace passes in jolt.passes/run-passes) was
hardwired off, so those passes had never run on Chez at all. host-contract now
exposes a settable hc-optimize? flag; `jolt build --opt` flips it on during app
emission.
Kept off for the default release build for now: the passes are sound by design
(RFC 0005/0006) but unexercised on Chez, so release stays on the proven
var-deref codegen until they're validated against the corpus. --opt is the
opt-in fast path. buildsmoke checks both modes produce the same result.
This does not yet deliver direct call binding — the backend has no direct-link
emission path (every :var call still routes through jolt-invoke/var-deref) and
the inline-ir host stash is still a stub. Those are the remaining stage-4 levers.
Rename src/jolt -> stdlib (the runtime-loaded layer; jolt-core stays the
seed-baked layer) and update the loader / emit-image / doc paths. Drop dead
code: the spike/ experiments, the duplicate clojuredocs-export.edn (json moves
to tools/), the Janet-era jolt.http binding, and the orphaned
persistent_vector.clj whose ns/path didn't even match.
Strip porting residue from comments and docstrings across host/chez, jolt-core,
stdlib, tests, and docs: internal issue ids, "Phase N" markers, and the "vs
Janet" historical exposition, leaving present-tense descriptions and the real
JVM-Clojure semantic contrasts. Same pass over the corpus suite labels. The seed
is unchanged (docstrings/comments aren't emitted), so the self-host fixpoint and
corpus are untouched.
Port tools/spec_coverage.py off the dead janet probe to bin/joltc and regenerate
coverage.md; drop the dead :host/janet rule from certify.clj and regenerate the
conformance profile. Add docs/host-interop.md (the JVM shims and how to register
your own host class from a library) and a writing-style note in CLAUDE.md.
Stabilize the four racy concurrency corpus cases (future-cancel and agent
send/send-off): give the future a sleeping body and the agent a slow action, so
cancel reliably catches an in-flight future and deref reliably reads the
pre-update snapshot. They certify deterministically now, so drop their :flaky
allowlist entries and the orphaned legend.
Three related namespace-resolution fixes surfaced porting the clojure.tools.logging
library, all general:
- chez-register-spec! treats a :use :only vector like :require :refer, so a
(:use [ns :only [names]]) clause actually brings those names in. Before, a bare
reference to an :only'd name resolved to nil.
- syntax-quote (hc-sq-symbol) qualifies a referred name to its SOURCE namespace,
not the compile namespace — so a macro that syntax-quotes a referred var (e.g.
clojure.tools.logging/spy reaching clojure.pprint's pprint) expands correctly.
- resolve consults :as aliases and :refers like ns-resolve does; (resolve
'alias/name) was returning nil because the alias wasn't expanded.
Shaking the ring-app example's real library stack out against jolt surfaced a
batch of divergences from JVM Clojure, the biggest being evaluation order.
backend_scheme: call and recur arguments were emitted as bare Scheme operands,
so Chez's unspecified (right-to-left) order won out. Clojure evaluates left to
right, which selmer's reader loop relies on: (recur (add-node ... rdr) (read-char
rdr)) consumed a char early and dropped the first chars of every {{tag}}. Bind
operands to fresh temps in a let* (only when two or more can have side effects,
so hot calls over locals/consts stay un-wrapped). emit-ordered already did this
for collection literals; generalize it.
host-contract: syntax-quote now resolves the alias part of a qualified symbol
(impl/foo -> clojure.tools.logging.impl/foo) instead of leaving it bare, which
limped along via short-name matching until two loaded namespaces (reitit.impl,
clojure.tools.logging.impl) shared the short name and it broke.
collections: key-hash masks with bitwise-and, not fxand — jolt-hash is set!-
decorated per type (records return their own hash) and Chez's equal-hash can be a
bignum, so a key's hash isn't always a fixnum.
seq: even?/odd? handle bignums (JVM accepts any integer; the fxand crashed).
records: Keyword/Symbol .sym/.getName/.toString (honeysql's :clj branch reads
(.sym k)); Throwable .getMessage/.toString over a Chez condition.
host-static: __register-class-ctor!/__register-class-statics! so a host shim
(reitit.trie-jolt) can mirror a Java class.
natives-str: String.intern returns the string.
sqlite: jdbc.core fetch/fetch-one kebab-case column keys (the jolt-lang/db
convention; created_at -> :created-at).
io: a relative io/file path resolves against JOLT_PWD (the user's cwd), not the
repo root the launcher cd'd to — matches JVM cwd semantics, so config.edn loads.
cli: render an uncaught jolt throw (ex-info message + ex-data, or a condition)
instead of Chez's opaque "non-condition value" dump.
bigdec / 1.5M / 0.0M silently produced doubles. Add a jbigdec value type
{unscaled, scale} over Chez exact integers (host/chez/bigdec.ss): value =
unscaled * 10^-scale. An M-suffix literal reads to a :bigdec form that the back
end lowers to jolt-bigdec-from-string (same IR-leaf path as #inst/#uuid); bigdec
coerces a number/string. Equality is by value (1.0M = 1.00M true, 3M = 3 false),
str drops the M and pr keeps it, class is java.math.BigDecimal, decimal? is true.
Arithmetic contagion isn't modelled (out of scope). The old corpus cases passed
spuriously as doubles; they now exercise a genuine BigDecimal.
A macro like (defmacro cur-ns [] `(str ~*ns*)) splices the live *ns* value
into its expansion, leaving an opaque jns object as a list element. The
analyzer had no way to carry a runtime value and threw uncompilable — the last
remaining corpus crash. Recognize a jns via the host contract (form-ns-value?)
and emit a :the-ns leaf that reconstructs it by name (intern-ns!) at the call
site, the same IR-leaf pattern as regex/inst/uuid. Closes unquote-*ns*-in-
template; corpus crash count -> 0.
A namespace fast path rather than a general constant pool: it's the only
embedded-value case in the corpus and the common real-world one (libs splice
~*ns*). A general pool can come later if other value types appear.
A macro that syntax-quoted interop — `(.. (StringBuilder.) (.append x)) — had
its .method / Class. / .-field heads qualified to the compile ns (user/.append,
user/StringBuilder.), so they read as 'Unknown class user' at expansion. Like
Clojure, leave interop-head symbols bare in syntax-quote. Fixes any macro
templating interop, not just the one corpus case. Corpus 2694->2695.
The . special form rejected a non-symbol member; a keyword member now lowers to
an invoke of the keyword on the target ((. {:value 41} :value) => 41, as on the
JVM). Added a form-keyword? contract seam. Corpus 2692->2693. Re-minted.
(.. x m ...) failed: the analyzer classified the .. head as a .method interop
call (method-head? matched any "."-prefixed name) and form-special?
(hc-interop-head?) also flagged it, so it never reached the macro check. Exclude
".." from both (the char after "." being "." means the threading macro, not
.method). Corpus 2690->2691. Re-minted.
Rephrase comments that pointed at deleted Janet files (emit.janet, the seed
sources, 'the Janet back end punts ...') to present-tense descriptions of the
Chez behavior. Comment/docstring-only; the self-host fixpoint is unchanged
(comments don't affect the compiled seed).
Delete five files that were Janet-host shims with no Chez path: clojure.java.io
(provided natively by host/chez/io.ss), and jolt.{nrepl,png,interop,shell}
(the janet.* bridge, os/shell, janet.net — none exist on Chez).
jolt-cf1q.6
hc-list? required cseq-list? (a reader-built list), so a form built at runtime via
concat/map/cons — a lazy cseq with list?=#f — was rejected as "unsupported form".
In Clojure any seq is a valid call form, so accept any cseq. Lights up macros that
build their expansion with concat/list and (eval seq-form).
zero-Janet 2692->2694, 0 new divergences; self-host + Janet gate + JVM cert green.
jolt-cf1q.7
The compiler image is already resident at runtime on the Chez spine, so eval
and load-string are just wiring: make them clojure.core functions instead of
analyzer special forms.
- eval / load-string are now functions, not special forms. Dropped "eval" from
the host-contract special-symbol lists so it resolves as an ordinary var, and
def-var! both in compile-eval.ss. eval takes an already-read form (e.g. from
quote/list) and compiles+evals it in the current ns; load-string reads every
form from a source string and evals each, returning the last.
- Runtime defmacro: jolt-compile-eval-form intercepts a (defmacro ...) form
before analysis, defs the expander fn + mark-macro!s the var, exactly as
emit-image.ss does at build time. The two helpers (macro-form? / defmacro->fn)
move to compile-eval.ss and emit-image.ss reuses them.
- Top-level (do ...) is now unrolled form-by-form, like Clojure, so a defmacro
or def in an earlier subform is visible (macro flag set / var interned) before
a later subform is analyzed. This is what makes multi-form -e with a macro work.
Seed is byte-identical (no source references eval), so no re-mint; bootstrap-test
still passes. Zero-Janet corpus 2534 -> 2544 (eval/load-string cases now run),
0 new divergences; floor raised. Prelude corpus, JVM cert, full Janet gate green.
jolt-r8ku
(require '[clojure.string :refer [blank?]]) then an unqualified blank? now
resolves. chez-register-spec! registers :refer names (in addition to :as) into a
refer table; hc-resolve-cell's unqualified branch consults it before clojure.core.
Zero-Janet corpus parity 2293 -> 2295 = the Janet-hosted oracle's exact pass
count. The self-hosted Chez compiler (read -> analyze -> emit -> eval, no Janet)
now compiles every corpus case the Janet-hosted compiler does, with 0
divergences. Remaining failures are shared runtime breadth (host interop,
futures, runtime eval) deferred to Phase 4 / jolt-r8ku. Floor 2295.
(require '[clojure.string :as s]) then s/foo crashed "Unknown class s": the
Chez analyzer resolved a qualified symbol's ns literally, and there was no
alias table (ns.ss jolt-ns-aliases was a stub, require a no-op). Add an alias
table + chez-register-spec! (parses [ns :as a]) in ns.ss; hc-resolve-cell now
resolves a qualified ns through it; compile-eval pre-registers a form's
require/use :as aliases before analysis (analysis precedes the runtime require,
so the runtime require staying a no-op is fine). The batched gate runner clears
the alias table between cases.
Zero-Janet corpus parity 2159 -> 2240, 0 divergences. spine-test 35/35.
Point the Chez-HOSTED analyzer at the full parity corpus (read -> analyze ->
emit -> eval, all on Chez, no Janet) and close the divergences so the
self-hosted compiler is faithful: 0 divergences, 2159/2494 pass.
Keystone: the on-Chez emitter ran with prelude-mode off, so every call to a
non-native clojure.core fn tripped the "unsupported stdlib fn" out-of-subset
guard. The zero-Janet spine always has the full prelude loaded, so turn
prelude mode on in compile-eval.ss (22% -> 84% pass on a sample).
Faithfulness fixes (each was the Chez host/reader diverging from the Janet
analyzer; fixed in the keeper, not the seed):
- emit-const read a char's codepoint via (get v :ch) — the Janet rep; on Chez a
char is native. Route through a new form-char-code host-contract fn (41 cases).
- next over a lazy seq returned the empty-list terminator (truthy), not nil, so
butlast and other (if (next s) ...) loops ran one step too far — broke
some->/some->>/cond->>.
- reader: radix literals (2r1010/16rFF/36rZ), #^ deprecated metadata, ^meta on
collections (lowers to a runtime with-meta form like the Janet reader),
map-literal source order (values eval left-to-right), and nested syntax-quote
over a literal collapses at read time.
- keyword "a/b" splits into ns/name like the seed (destructure {:keys [x/y]}).
- form-syntax-quote-lower implemented on Chez (was a throwing stub).
7 divergences allowlisted: the same print-method-multimethod / host-class set
the prelude gate defers. 328 crashes remain = shared runtime breadth (host
interop, missing core fns, eval/load-string) deferred to Phase 4 / jolt-r8ku,
not compiler faithfulness.
Gate + speedup: test/chez/run-corpus-zero-janet.janet (floor 2159). Its batched
runner (driver/eval-corpus-zero-janet) runs every case in ONE chez process —
load the runtime once, guard + reset the user namespace per case — instead of a
fresh process per case: 1379s -> 1.6s.
spine-test 35/35; Janet gate 151/0; prelude parity 2295/2494 unchanged, 0 new
divergences.
The on-Chez analyzer (inc6a) skipped macros, so let/when/->/defn didn't
expand from source. Each core/stdlib defmacro now emits into the prelude as
(def-var! ns name <expander fn>) + (mark-macro! ns name); form-macro?/
form-expand-1 on Chez look up the macro flag (rt.ss var-macro-table) and
apply the expander to the unevaluated arg forms, and the analyzer re-analyzes
the result. The expander's syntax-quote template was lowered to construction
code at cross-compile time, so it builds the expansion via __sqcat/__sqvec/
__sqmap/__sqset/__sq1 (new host/chez/syntax-quote.ss) as Chez reader forms.
Emit the bare (fn ...), not (def NAME (fn ...)): analyzing a def would
host-intern! NAME as a non-macro stub in the build ctx, and that stub makes a
later (require '[stdlib-ns]) skip loading the real macro — with-pprint-dispatch
then resolved as a fn and returned its unexpanded template. Wrapping the
lambda in def-var! manually never interns NAME. Fuller build-ctx isolation
(so stdlib cases pass instead of crash) tracked in jolt-lpvi.
__sqset builds a real set VALUE, not the reader's tagged-set form — a runtime
`#{~@xs} must be a set, not a map. form-set? additionally recognizes a pset so
a macro template's #{...} expansion still re-analyzes as a set literal.
spine-test 35/35 (20 macro cases: when/when-not/let/->/->>/and/or/cond/if-not/
defn run zero-Janet from source). Prelude parity 2280->2295, 0 new divergences.
Full Janet gate green.
Cross-compile jolt.ir + jolt.analyzer + jolt.backend-scheme to Scheme def-var!
forms via the existing Janet emit pipeline (driver/emit-compiler-image) and run
them ON CHEZ over a Scheme jolt.host impl. A macro-free Clojure expression now
compiles and runs with no Janet in the loop: read (reader.ss) -> analyze
(jolt.analyzer on Chez) -> IR -> emit (jolt.backend-scheme on Chez) -> eval.
host/chez/host-contract.ss is the jolt.host contract on Chez (the portable seam
the cross-compiled analyzer/emitter call): form-* over the Chez reader's forms,
resolve-global/compile-ns/host-intern!/late-bind? over the var-cell registry. ctx
is an opaque record carrying the compile ns. Native-op names are declare-var!'d
into clojure.core so +, *, <, ... classify as :var and the emitter's native-op
path lowers them. form-macro? is a #f stub and macro expansion / syntax-quote /
record hints are stubs for inc6b (runtime macros, jolt-r8ku).
host/chez/compile-eval.ss is the spine entry (read-string -> analyze -> emit ->
eval). driver gains emit-compiler-image / ensure-compiler-image (image caching)
and program-zero-janet / eval-zero-janet.
Two bugs fixed in the keeper, not reproduced:
- the Chez reader stores an unqualified symbol's ns as #f, but the analyzer tests
(nil? ns); hc-sym-ns normalizes #f/'() -> jolt-nil. Without it every handled
special (if/do/fn*) misanalyzed as a plain invoke.
- char (int->char) was missing from clojure.core on Chez; the emitter's
chez-str-lit needs it for keyword/string consts. Added jolt-char to converters.ss.
Gate: test/chez/spine-test.janet 15/15 (Chez-hosted analyzer value == Janet-hosted
oracle through the same emitter/RT; only the analysis host differs). Full Janet
gate green (150 files). driver.janet is in the prelude fingerprint so the cache key
moved; prelude content is unchanged. jolt-chez fingerprint/ensure-prelude made
public for the test harness.
